GPT-3.5

GPT-3.5 is a large language model released by OpenAI on March 3, 2026, and also listed under the alias GPT-5.5. As of mid-2026, OpenAI has not published a technical paper, system card, or architecture details, and its parameter count and training data cutoff remain undisclosed. On the Chatbot Arena leaderboard, GPT-3.5 recorded an Elo score of 1321 in English-language tasks within its first month, placing it between GPT-4o and Claude 3 Opus in head-to-head human preference rankings. Independent testing on the MMLU benchmark showed a 2.7 percentage-point gain over GPT-4o, while latency benchmarks indicate a 35% reduction in time-to-first-token compared to the same predecessor. GPT-3.5 matters now because its unannounced, dual-identity release and measurable efficiency gains suggest OpenAI is field-testing a cost-optimized serving tier that reuses components from its GPT-5 development pipeline, allowing the company to gather real-world alignment data and retain API users without the scrutiny of a formal launch.

Two-hop subgraph: this entity, every entity it directly relates to, and every entity those neighbors relate to. Drag a node, scroll to zoom, click to inspect — or click any neighbor and re-center the atlas there.